Why most AI pilots stall before production

Almost every stalled AI programme we are called into has a working model. The notebook runs, the numbers look defensible, and someone has already presented them to a steering committee. What is missing is everything between that notebook and a system that a business process can depend on.

Three gaps account for most of it. First, nobody owns the model in production, because the data science team is funded for discovery and the platform team never agreed to operate something they did not build. Second, there is no evaluation harness, so the only way to know the model still works is to wait for a complaint. Third, the deployment path does not exist: no CI, no rollback, no way to serve a prediction to the system that needs it.

None of those are modelling problems, which is why more modelling does not fix them. The cheapest correction is to decide who operates the thing before you build it, and to make the first workload deliberately small enough that the operating question has a real answer.

A useful test: ask who gets paged when the model degrades at two in the morning. If the honest answer is nobody, the pilot is not close to production regardless of how good the metrics look.